Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Benchmarking 3D multi-coil NC-PDNet MRI reconstruction

Asma Tanabene, Chaithya Giliyar Radhakrishna, Aurélien Massire, Mariappan S. Nadar, Philippe Ciuciu

TL;DR

This work extends the Non-Cartesian Primal-Dual Network to a 3D multi-coil setting, and shows that NC-PDNet trained on compressed data with varying input channel numbers achieves an average PSNR of 42.98dB for 1 mm isotropic 32 channel whole-brain 3D reconstruction.

Abstract

Deep learning has shown great promise for MRI reconstruction from undersampled data, yet there is a lack of research on validating its performance in 3D parallel imaging acquisitions with non-Cartesian undersampling. In addition, the artifacts and the resulting image quality depend on the under-sampling pattern. To address this uncharted territory, we extend the Non-Cartesian Primal-Dual Network (NC-PDNet), a state-of-the-art unrolled neural network, to a 3D multi-coil setting. We evaluated the impact of channel-specific versus channel-agnostic training configurations and examined the effect of coil compression. Finally, we benchmark four distinct non-Cartesian undersampling patterns, with an acceleration factor of six, using the publicly available Calgary-Campinas dataset. Our results show that NC-PDNet trained on compressed data with varying input channel numbers achieves an average PSNR of 42.98 dB for 1 mm isotropic 32 channel whole-brain 3D reconstruction. With an inference time of 4.95sec and a GPU memory usage of 5.49 GB, our approach demonstrates significant potential for clinical research application.

Benchmarking 3D multi-coil NC-PDNet MRI reconstruction

TL;DR

This work extends the Non-Cartesian Primal-Dual Network to a 3D multi-coil setting, and shows that NC-PDNet trained on compressed data with varying input channel numbers achieves an average PSNR of 42.98dB for 1 mm isotropic 32 channel whole-brain 3D reconstruction.

Abstract

Deep learning has shown great promise for MRI reconstruction from undersampled data, yet there is a lack of research on validating its performance in 3D parallel imaging acquisitions with non-Cartesian undersampling. In addition, the artifacts and the resulting image quality depend on the under-sampling pattern. To address this uncharted territory, we extend the Non-Cartesian Primal-Dual Network (NC-PDNet), a state-of-the-art unrolled neural network, to a 3D multi-coil setting. We evaluated the impact of channel-specific versus channel-agnostic training configurations and examined the effect of coil compression. Finally, we benchmark four distinct non-Cartesian undersampling patterns, with an acceleration factor of six, using the publicly available Calgary-Campinas dataset. Our results show that NC-PDNet trained on compressed data with varying input channel numbers achieves an average PSNR of 42.98 dB for 1 mm isotropic 32 channel whole-brain 3D reconstruction. With an inference time of 4.95sec and a GPU memory usage of 5.49 GB, our approach demonstrates significant potential for clinical research application.

Paper Structure

This paper contains 13 sections, 3 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Experimental Setup
  3. Model
  4. Data
  5. Non-Cartesian Sampling Patterns
  6. Implementation, training and evaluation details
  7. Results and Analysis
  8. Channel-Specific vs. Channel-Agnostic Training
  9. Performance benchmarking of selected non-Cartesian sampling patterns
  10. Assessing the Impact of Coil Compression
  11. DISCUSSION and CONCLUSION
  12. Compliance with ethical standards
  13. Acknowledgments

Figures (5)

  • Figure 1: NC-PDNet architecture ncpdnet for multi-coil MRI reconstruction, where $N_f$ is the number of convolutional filters, $N_P$ is the buffer size, and $N_C$ is the number of unrolled iterations.
  • Figure 2: Trajectories are shown with a reduced number of shots for clarity. GoLF-SPARKLING combines a non-Cartesian SPARKLING portion (blue) with a grid-sampled low-frequency region (green).
  • Figure 3: Quantitative results of NC-PDNet with different trajectories, ordered from left to right: TPI, 3D Radial, 3D Cones, GoLF-SPARKLING (GS w/CC), and GoLF-SPARKLING without coil compression (GS).
  • Figure 4: Reconstruction results of the $90^{th}$ slice of file e14079s3_P09216.7 from the test set. The top row shows reconstructions by different methods, while the bottom row displays zoomed-in regions outlined by red frames. Volume-wise PSNR and SSIM scores are indicated in the top left corner of each image.
  • Figure 5: Evolution of test scores over cumulative training time for channel-agnostic NC-PDNet training.